Thanks again for your feedback.

On Sat, Sep 10, 2016 at 9:55 PM, Stefan Monnier 
<monnier@iro.umontreal.ca> wrote:
> What'd be the benefit?  You do want to `defvar' them in order to give
> them a doctring (which means that if you want them void you have to do
> extra gymnastic), and giving them a good default value means you can
> call them without having to check the value beforehand (and that also
> means you can use `add-function' on it).

Nothing, but I couldn't think of any downsides either before you
mentioned them. I see now that a `defvar' is better.

> `buffer-hash` is only a problem in large buffers.  But you can use
> buffer-text-modified-tick to get a much quicker test (only important 
> in
> large buffers).  And you might not even need to test
> buffer-text-modified-tick because you can often just flush the 
> (relevant
> part of) the cache(s) from an after-change-functions or from
> syntax-propertize or ...

OK, I changed it to use the buffer's tick counter instead. I also
changed `seq-uniq' to `delete-dups', which resulted in a massive
speedup. Some benchmarks using the same test files as before follow.

* With `seq-uniq':
(benchmark 10 '(css--foreign-completions 'css-class-list-function))
    "Elapsed time: 4.198944s (0.911449s in 60 GCs)"

* With `delete-dups':
(benchmark 10 '(css--foreign-completions 'css-class-list-function))
    "Elapsed time: 0.282890s (0.188205s in 10 GCs)"

As a side note, maybe a hashing strategy like the one `delete-dups'
uses would be good for `seq-uniq' too?

> You don't need the boundp test here, AFAICT.

Right, removed.

-- Simen